How far behind are Chinese AI models, and does "cheap and good enough" win?
Let's start with the uncomfortable number. On the new GDPx agent benchmark, the leader scores 12%, and the best open-weight Chinese-built model in the table, GLM 5.3, scores 2.7%. On the toughest long tasks, that is not a small gap.
Now the plot twist. DeepSeek has been putting its energy somewhere else. It released its V4-Pro model officially with higher API prices, put out an open-source agent harness, and on 31 October published a coding super-agent under the MIT license that picked up around 4,000 GitHub stars within a day. It is pitched as a rival to Claude Code.
So what is the game here?
Look at the pattern. Raising prices while giving away the tooling is not what a pure price-war company does. It looks like a lab moving from "cheap model" to "the place where developers build agents".
And the question underneath all of it: if a model is good enough and costs a fraction, why would anyone pay for the best one? Honestly, it depends on the job. For a task where 12% is the ceiling, "good enough" is a moving target.
One more thing worth being straight about: there is no market share data in any of this. What we can see is where the gap is large (hard, long tasks), where it seems to be closing (tools and distribution), and that the price advantage is not a fixed law.
The cheap model wins when the work is routine. The best model wins when a mistake is expensive. The real fight is over where that line sits, and it moves every few months.